Testing machine learning (ML) models for robustness and accuracy using generative deep learning

Abstract

A model testing system administers tests to machine learning (ML) models to test the accuracy and the robustness of the ML models. A user interface (UI) associated with the model testing system receives selections of one or more of a plurality of tests to be administered to a ML model under test. Test data produced by one or more of a plurality of testing ML models that correspond to the plurality of tests is provided to the ML model under test based on the selected tests. One or more of a generative patches test, a generative perturbations test and a counterfeit data test can be administered to the ML model under test based on the selections.

What is claimed is: 
     
       1. A machine learning (ML) based model testing system comprising:
 at least one processor; 
 a non-transitory processor readable medium storing machine-readable instructions that cause the processor to:
 receive a selection of one or more tests to be administered to a ML model under test, wherein the one or more tests include at least one of a generative patches test, a generative perturbations test and a counterfeit data test; 
 access a source database that stores training data with data samples that are used for training the ML model under test; 
 provide at least a subset of the data samples from the source database to one or more of a plurality of testing ML models that correspond to the selected one or more tests; 
 obtain modified data samples from the one or more testing ML models that correspond to the selected one or more tests, wherein the modified data samples are produced by the one or more testing ML models from the subset of data samples; 
 execute each of the selected one or more tests on the ML model under test using the modified data samples; and 
 generate a report that includes results of the selected one or more tests, the results conveying a successful test or an unsuccessful test for each of the selected one or more tests, wherein the successful test pertains to the ML model under test producing accurate results and the unsuccessful test pertains to the ML model under test producing inaccurate results. 
 
 
     
     
       2. The ML model testing system of  claim 1 , further comprising instructions that cause the processor is to:
 train the plurality of testing ML models for generating the modified data samples using unsupervised learning. 
 
     
     
       3. The ML model testing system of  claim 1 , wherein to receive the selection of one or more tests to be administered, the processor is to:
 provide a user interface (UI) that presents options for the selection of the one or more tests to be administered to the ML model under test. 
 
     
     
       4. The ML model testing system of  claim 1 , further comprising instructions that cause the processor is to:
 store the modified data samples from each of the plurality of ML testing models to a respective test datastore. 
 
     
     
       5. The ML model testing system of  claim 4 , wherein to execute each of the selected one or more tests on the ML model under test, the processor is to:
 select the modified data samples to provide to the ML model under test from one or more of the respective test datastores based on the selected one or more tests to be administered. 
 
     
     
       6. The ML model testing system of  claim 1 , wherein the ML model under test includes one of an image classifier and an image search model and the data samples include images that are used for training the ML model under test. 
     
     
       7. The ML model testing system of  claim 6 , wherein each of the plurality of testing ML models includes a generative adversarial network (GAN). 
     
     
       8. The ML model testing system of  claim 7 , wherein each of the testing ML models includes at least a generator and a discriminator. 
     
     
       9. The ML model testing system of  claim 7 , wherein the one or more testing ML models include at least a patch generation model and to obtain the modified data samples from the patch generation model during the generative patches test, the processor is to:
 obtain as output from the patch generation model, the subset of data samples modified with patches that include modifications to one or more of temporal and spatial features of the subset of data samples. 
 
     
     
       10. The ML model testing system of  claim 7 , wherein the one or more testing ML models include at least a perturbations generation model and to obtain the modified data samples from the perturbations generation model during the generative perturbations test, the processor is to:
 obtain as output from the perturbation generation model, the subset of data samples modified with perturbation that include modifications to one or more pixels of the images in the subset of data samples. 
 
     
     
       11. The ML model testing system of  claim 7 , wherein the one or more testing ML models include at least a counterfeit data generation model and to obtain the modified data samples from the counterfeit data generation model during the counterfeit data test, the processor is to:
 obtain as output from the counterfeit data generation model, counterfeit images that are generated based on the images in the subset of data samples. 
 
     
     
       12. A method comprising:
 training a plurality of testing machine learning (ML) models to produce corresponding modified data samples from a source database that stores training data with data samples that are used for training a ML model under test, the plurality of testing ML models corresponding to a plurality of tests to be administered to the ML Model under test; 
 receiving a selection of one or more of the plurality of tests to be administered to the ML model under test; 
 providing at least a subset of the data samples from the source database to one or more of the plurality of testing ML models that correspond to the selected one or more tests; 
 executing the selected one or more tests on the ML model under test using modified data samples that are obtained by providing the subset of the data samples to the one or more testing ML models that correspond to the selected one or more tests; and 
 generating a model report with results of the selected one or more tests administered to the ML model under test, the results including indications for successful test results wherein the ML model under test produces accurate results and unsuccessful test results wherein the ML model under test produces inaccurate results. 
 
     
     
       13. The method of  claim 12 , wherein the modified data samples include one or more images with patches and perturbations and the ML model under test includes an object recognition ML model and the successful test includes the object recognition ML model identifying previously-viewed objects from the modified data samples. 
     
     
       14. The method of  claim 12 , wherein the modified data samples include counterfeit images generated by one of the plurality of testing ML models and genuine images from the source database and the ML model under test includes an image search ML model and a successful test result includes the image search ML model identifying the counterfeit images. 
     
     
       15. The method of  claim 12 , further comprising:
 obtaining the modified data samples by providing the data samples from the source database to the one or more testing ML models. 
 
     
     
       16. The method of  claim 15 , further comprising:
 modifying one or more of spatial and temporal data of images included in the subset of data samples by a generator of a patch generation model included in the plurality testing ML models wherein each of the plurality of testing ML models includes a generative adversarial network (GAN). 
 
     
     
       17. The method of  claim 16 , further comprising:
 providing the modified images to a discriminator included in the patch generation model; 
 identifying one or more of the modified images that the discriminator fails to identify as including the modified spatial and temporal data; and 
 providing to the ML model under test, the one or more of the modified images that the discriminator fails to identify as modified images. 
 
     
     
       18. A non-transitory processor-readable storage medium comprising machine-readable instructions that cause a processor to:
 receive a selection of one or more tests to be administered to a ML model under test, wherein the one or more tests include at least one of a generative patches test, a generative perturbations test and a counterfeit data test; 
 access a source database that stores training data with data samples that are used for training the ML model under test; 
 provide at least a subset of the data samples from the source database to one or more of a plurality of testing ML models that correspond to the selected one or more tests; 
 obtain modified data samples from the one or more testing ML models that correspond to the selected one or more tests, wherein the modified data samples are produced by the one or more testing ML models from the subset of data samples; 
 execute each of the selected one or more tests on the ML model under test using the modified data samples; and 
 generate a report that includes results of the executed one or more tests, the results conveying a successful test or an unsuccessful test for each of the selected one or more tests, wherein the successful test pertains to the ML model under test producing accurate results and the unsuccessful test pertains to the ML model under test producing inaccurate results. 
 
     
     
       19. The non-transitory processor-readable storage medium of  claim 18 , wherein each of the plurality of testing ML models includes a generative adversarial network (GAN) that includes at least a generator and a discriminator and the data samples includes images. 
     
     
       20. The non-transitory processor-readable storage medium of  claim 19 , further comprising instructions that cause the processor to:
 provide the images with modified pixel data as to a discriminator included in a perturbation generation model that forms one of the plurality of testing ML models; 
 identify one or more of the modified images that the discriminator fails to identify as including the modified pixel data; and 
 provide to the ML model under test as test data, the one or more of the modified images that the discriminator fails to identify as the modified data samples.